RR-TB Updates - Aug. 31

Updates

• Updated HIV values (to prioritize test result, then self-report)
• Ran models for relapsed and re-entry cases; Played with K settings in smooths (effect is pretty marginal)
• Health Unit:
  • Not added to models yet
  • I noticed there was a lot of missingness in Sao Paulo between 2018-2019 (it looks like it jumps from a few hundred missing cases to >20k, and then back down in 2020) – I’m double checking whether this is an issue on my end or worth asking the Brazil team about
• Finalized list of potential individual and municipal-level covariates to add in and found data sources:
  • Urbanicity, defined as proportion of municipality pop living in urban setting
    • Source: IBGE - 2010 census
  • Bolsa Familia Coverage
    • Source: Baroni et al., 2021 paper - Database on the coverage of the “Bolsa-Família” conditioning cash-transfer program: Brazil, 2005 to 2021)
  • FHS Coverage, defined as number of people per municipality registered in 2015
    • Source: DATASUS
  • Household crowding, Defined as percent of household with > 2 people living in a sleeping area
    • Source: 2010 Census, via Atlas BR
  • Treatment abandonment rate per municipality in period quarter (or year)
    • Source: Calculate from Sinan

New Cases:

• I pooled the estimated effects and standard errors across groups for each quarter in each state for the models below - they look a little off in some places, so I need to go back and check them out.

Model 1: Random intercept

gam(cbind(Positive, Negative) ~ s(state, bs = "re", k = 27) + s(time, bs = "tp", k = 10) 
    + age_cat + hiv_status + sex, 
             data = mdf_new_grp, 
             family = binomial (link = "logit"), 
             method = "REML")

Model 2: Random slope

gam(cbind(Positive, Negative) ~ s(time, bs = "tp", k = 10) + s(time, state, bs = "re", k = 27) 
    + age_cat + hiv_status + sex, 
                           data = mdf_new_grp, 
                           family = binomial (link = "logit"), 
                           method = "REML")

Model 3: Random intercept + Random slope

gam(cbind(Positive, Negative) ~ s(state, bs = "re", k = 27) + s(time, state, bs = "re", k = 27) 
    + age_cat + hiv_status + sex, 
              data = mdf_new_grp, 
              family = binomial (link = "logit"), 
              method = "REML")

Note: For the 95% CIs, the Y axes vary for each state to capture the state-specific variation over time.

Same plot as above, just with fixed Y-axis.

Model 4: GI - Random intercept + group-level smooths with varying smoothness

gam(cbind(Positive, Negative) ~ s(state, bs = "re", k=27) + s(time, bs = "tp", k = 10) 
    + s(time, by = state, bs = "tp", m=1, k = 10) + age_cat + hiv_status + sex, 
                           data = mdf_new_grp, 
                           family = binomial (link = "logit"), 
                           method = "REML")

Model 5: GS - Global smooth + group-level smooths with same smoothness

gam(cbind(Positive, Negative) ~ s(time, bs = "tp", m=2) 
    + s(time, state, bs = "fs", m = 2) + age_cat + hiv_status + sex, 
              data = mdf_new_grp, 
              family = binomial (link = "logit"), 
              method = "REML")

Compare models

I also fit the above without the number tested to get a better sense of how the modeled trends compare to each other

Re-Entry:

Combined

I figured I would present these all together since each model is still noisy, especially in states where there are no rif cases (e.g. Acre, Roraima, and Tocantins). I quickly looked into potential inclusion criteria for percent tested, which would help address some of these issues, especially we define the inclusion criteria as consistently testing > 5% of case population for at least two consecutive periods. This would largely exclude 2014-2015. I think we would want to add a threshold to the minimum number of cases tested, since some state who test a few folks, but who have only a few cases overall would exceed the threshold.

Relapse

Combined: